Research On Question Generation Oriented Exposure Bias Resolution And Multi-object Evaluation Method

Question Generation(abbr.,QG)is an important task in the field of Question Answering.It aims to automatically generate grammatically smooth and semantically related natural questions according to a given text.This technology can bring machines the ability to ask questions independently,and has a wide application prospect in emerging intelligent education,automatic inquiry and community question answering.The existing research often adopts the way of teacher-forcing learning,which drives the QG model to generate questions that are similar or even consistent with the reference questions.However,the QG models have no reference questions to refer to in the tests,but iteratively predict the following questions based on the unconvinced information that previously predicted.The different information between training and testing that known as exposure bias,leads to the accumulation of errors in the generation process of the model.This paper carries out the following specific research on the above problems:1)The essence of exposure bias refers to the difference of reliable context information.There is a large am ount of reliable information in training and a small amount of reliable information in tests.Therefore,this paper proposes a random noise-aware training method,which mixes a certain proportion of random noise into reference questions to reduce the amount of reliable information in the training stage.It enhances the model’s perception of error information,so that the QG model can fight against accidental errors during tests.2)In addition,this paper finds that the errors produced by the model during tests often have a specific semantic correlation with its context.Based on this,this paper proposes a readable noise-aware training method,which uses the mask language model to produce noise that is more in line with grammar and context.Compared with random noise,readable noise is more similar to the actual errors in the test stage,which can fit the occasional errors in the test environment and further alleviate the problem of exposure bias.3)The above research promotes the QG models against exposure bias.However,it also restricts the flexibility of the model.The QG models entirely adhere to a standardized generation mode,pursue the unique and definite reference question,which reduces the diversity of generated questions.To alleviate this problem,this paper proposes a multi-object evaluation method for QG and constructs the corresponding diversified evaluation data.Furthermore,this method evaluates the quality of the generated questions according to multiple reference questions.Thus,the relationship between exposure bias and model flexibility can be objectively reflected.The two noise-aware training methods proposed in this paper alleviate the exposure bias of QG model.The experiment results on the SQuAD dataset show that our method achieved 22.31%of BLEU-4.In addition,the multi-object QG evaluation method enriches the research perspective and verification methods of exposure bias resolution.